1
00:00:00,000 --> 00:00:09,960
Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the

2
00:00:09,960 --> 00:00:15,700
latest trends, tips, and tools in the world of marketing AI, helping you get the best

3
00:00:15,700 --> 00:00:18,520
results from your marketing efforts.

4
00:00:18,520 --> 00:00:23,160
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:23,160 --> 00:00:25,080
Hello everybody, it's Paul Avery here.

6
00:00:25,080 --> 00:00:29,240
Welcome to episode 39 of Artificially Intelligent Marketing.

7
00:00:29,240 --> 00:00:33,160
I'm here with my co-host and the lovely Martin Broadhurst.

8
00:00:33,160 --> 00:00:34,400
How are we Martin?

9
00:00:34,400 --> 00:00:36,960
I'm splendid.

10
00:00:36,960 --> 00:00:37,960
Just splendid.

11
00:00:37,960 --> 00:00:38,960
That's it.

12
00:00:38,960 --> 00:00:41,000
A single word will do it.

13
00:00:41,000 --> 00:00:42,000
I'm not splendid.

14
00:00:42,000 --> 00:00:47,680
This has got nothing to do with marketing or AI, but we do talk about football or soccer

15
00:00:47,680 --> 00:00:48,880
on the podcast sometimes.

16
00:00:48,880 --> 00:00:54,920
Usually it's me teasing Martin about just how useless Derby County are, but I'm a Liverpool

17
00:00:54,920 --> 00:01:00,120
fan and today our talisman, Jurgen Klopp, has announced he'll be leaving the club.

18
00:01:00,120 --> 00:01:04,480
So I'm really going to have to muster energy today, Martin, for this conversation because

19
00:01:04,480 --> 00:01:05,840
I'm really sad.

20
00:01:05,840 --> 00:01:08,240
Yeah, it was a big announcement.

21
00:01:08,240 --> 00:01:11,480
I was shocked, but I could see why it might impact you.

22
00:01:11,480 --> 00:01:17,480
There is going to be a grieving process to grow through, to go and grow through.

23
00:01:17,480 --> 00:01:23,040
So yeah, I accept there's going to be denial, anger, and at some point you'll land on acceptance.

24
00:01:23,040 --> 00:01:25,880
Hopefully I can get to that quickly.

25
00:01:25,880 --> 00:01:30,360
Fortunately, as much as that's big news, there's been more big news in the world of marketing

26
00:01:30,360 --> 00:01:31,880
and AI.

27
00:01:31,880 --> 00:01:33,960
We've got some interesting stuff to talk through today.

28
00:01:33,960 --> 00:01:39,440
We've got the UK government publishing some more guidance on generative AI from employees.

29
00:01:39,440 --> 00:01:40,680
We're going to dive into it.

30
00:01:40,680 --> 00:01:43,000
I think that's going to be really interesting.

31
00:01:43,000 --> 00:01:47,440
We've got Google releasing a new text-to-video model that none of us can play with yet, but

32
00:01:47,440 --> 00:01:49,640
as usual, the demo videos look awesome.

33
00:01:49,640 --> 00:01:54,320
So this could be the next improvement as we iterate our way up towards production-level

34
00:01:54,320 --> 00:01:57,160
video quality from generative AI.

35
00:01:57,160 --> 00:02:01,120
Runway, one of our favorite companies, as listeners will know, we're going to talk about

36
00:02:01,120 --> 00:02:03,600
what they've been doing to improve their video tool.

37
00:02:03,600 --> 00:02:10,680
We've got some new enhancements and cost reductions coming to OpenAI's platform for those that

38
00:02:10,680 --> 00:02:12,680
use the API.

39
00:02:12,680 --> 00:02:18,000
Eleven Labs, Voice Synthesis fame have got a ton of cash to improve their product even

40
00:02:18,000 --> 00:02:19,000
further.

41
00:02:19,000 --> 00:02:25,800
We're going to talk about the ongoing evolution of AI website builders and is it all hype

42
00:02:25,800 --> 00:02:27,440
or are they actually any good?

43
00:02:27,440 --> 00:02:29,280
We're going to talk about that as well.

44
00:02:29,280 --> 00:02:32,200
We've got an MIT study on whether or not AI is going to replace us all.

45
00:02:32,200 --> 00:02:33,600
That's going to be quite interesting.

46
00:02:33,600 --> 00:02:38,280
Obviously, we've got, for those of you that watch TV, you'll have seen the ads for Samsung

47
00:02:38,280 --> 00:02:46,360
Galaxy S24 with its AI baked in and the rise of AI phones and then a bunch of other stuff

48
00:02:46,360 --> 00:02:47,360
all in between.

49
00:02:47,360 --> 00:02:51,280
I think, Myon, between us, we're going to help me get over the Yergin Klopp news by

50
00:02:51,280 --> 00:02:54,160
looking at all this exciting AI news instead.

51
00:02:54,160 --> 00:02:55,640
What do you reckon?

52
00:02:55,640 --> 00:03:00,120
Well, there's plenty to go at and if nothing else, it'll be a short-term distraction.

53
00:03:00,120 --> 00:03:01,720
Well, let's do it.

54
00:03:01,720 --> 00:03:02,720
Let's get stuck in.

55
00:03:02,720 --> 00:03:07,360
Let's talk about this new guidance from the UK government, Myon.

56
00:03:07,360 --> 00:03:15,520
They've published a generative AI framework for HM Government, which is a 70-odd page

57
00:03:15,520 --> 00:03:23,920
document covering all things about safe, responsible, and effective use of generative AI in government

58
00:03:23,920 --> 00:03:26,160
and the public sector.

59
00:03:26,160 --> 00:03:31,920
I've been fairly critical of this government and certain episodes in the past, I've given

60
00:03:31,920 --> 00:03:36,160
them a bit of a bashing, but this document is very good.

61
00:03:36,160 --> 00:03:37,160
I applaud them.

62
00:03:37,160 --> 00:03:44,080
I think they've done a really comprehensive job when it comes to explaining what AI is

63
00:03:44,080 --> 00:03:52,120
and generative AI, explaining how to use it, giving some examples, talking about how to

64
00:03:52,120 --> 00:03:57,080
procure it, talking about the ethics and responsible use.

65
00:03:57,080 --> 00:04:02,120
It's a very broad document, but the language that they write in is plain English.

66
00:04:02,120 --> 00:04:06,080
It's technical without being overly technical.

67
00:04:06,080 --> 00:04:13,240
If you're just a departmental head in a local authority, you could pick up this and read

68
00:04:13,240 --> 00:04:20,240
it without needing to know anything about deep learning and neural networks and all

69
00:04:20,240 --> 00:04:21,240
of that kind of thing.

70
00:04:21,240 --> 00:04:23,920
They've done a really good job.

71
00:04:23,920 --> 00:04:34,520
It's based around 10 core principles and they're very easy to understand.

72
00:04:34,520 --> 00:04:36,360
There's no complexity to it.

73
00:04:36,360 --> 00:04:41,560
Now, the 10 core principles, I'll give you a summary of what they include.

74
00:04:41,560 --> 00:04:48,840
It talks about you should be operating with an understanding of what generative AI is

75
00:04:48,840 --> 00:04:53,760
and its limitations, recognizing what it can and can't do.

76
00:04:53,760 --> 00:05:02,220
You should understand what it is to use generative AI in a lawful, ethical, and responsible way.

77
00:05:02,220 --> 00:05:11,000
They talk about security, data privacy, implementing technical controls about who can access the

78
00:05:11,000 --> 00:05:12,000
data.

79
00:05:12,000 --> 00:05:17,240
They talk about meaningful human control, which is about emphasizing the necessity of

80
00:05:17,240 --> 00:05:18,240
human oversight.

81
00:05:18,240 --> 00:05:22,600
They're kind of human in the loop, the kind of thing that we talk about all the time on

82
00:05:22,600 --> 00:05:27,080
the podcast when it comes to using these systems.

83
00:05:27,080 --> 00:05:28,080
There's more as well.

84
00:05:28,080 --> 00:05:33,240
I'm not going to go into all of them, but they very much talk about skills and expertise,

85
00:05:33,240 --> 00:05:37,840
being collaborative with commercial colleagues from the start.

86
00:05:37,840 --> 00:05:44,080
I think the thing that really stood out to me is that they're really taking generative

87
00:05:44,080 --> 00:05:52,920
AI and the potential benefits as well as the potential downsides and risks.

88
00:05:52,920 --> 00:05:54,520
They're taking it very seriously.

89
00:05:54,520 --> 00:05:56,440
They're not shying away from this.

90
00:05:56,440 --> 00:05:59,600
This is very much an innovation first approach from government.

91
00:05:59,600 --> 00:06:02,280
I'm very keen to see it.

92
00:06:02,280 --> 00:06:09,880
The thing that really came through for me in the forward and in the executive summary

93
00:06:09,880 --> 00:06:15,080
is they're really pushing AI literacy.

94
00:06:15,080 --> 00:06:19,840
This is big to the point of, well, you can see this as a big change management piece

95
00:06:19,840 --> 00:06:26,960
for them as evidenced by additional resources outside of this PDF being made available.

96
00:06:26,960 --> 00:06:34,720
One such document is a poster that lists the 10 principles and the 10 core principles of

97
00:06:34,720 --> 00:06:36,600
using generative AI.

98
00:06:36,600 --> 00:06:43,000
They're encouraging departments and managers to access these posters and put them up around

99
00:06:43,000 --> 00:06:44,440
their workplace.

100
00:06:44,440 --> 00:06:51,800
You're going to start seeing local councils and all these government bodies having posters

101
00:06:51,800 --> 00:06:55,020
talking about best practice for using generative AI.

102
00:06:55,020 --> 00:06:56,640
They are not shying away from it.

103
00:06:56,640 --> 00:06:58,680
They're really, really leaning into it.

104
00:06:58,680 --> 00:07:06,520
Is this document, this is for internal use within county councils and government institutions?

105
00:07:06,520 --> 00:07:11,600
This is not more broad advice for small businesses, for example?

106
00:07:11,600 --> 00:07:20,440
No, this is very much for internal use within government, local government, NHS, all these

107
00:07:20,440 --> 00:07:22,240
government departments.

108
00:07:22,240 --> 00:07:25,640
It's about how to use this in a safe and responsible way.

109
00:07:25,640 --> 00:07:27,600
However, it applies to everyone.

110
00:07:27,600 --> 00:07:33,120
You could pick up this document and do a find and replace government for my business and

111
00:07:33,120 --> 00:07:35,040
it would be just as valid.

112
00:07:35,040 --> 00:07:40,920
It's a really well crafted document and I think it's plain English to the point where

113
00:07:40,920 --> 00:07:44,040
anyone can pick it up and learn from it.

114
00:07:44,040 --> 00:07:50,760
I think that's important because the world of generative AI, the technical aspects become

115
00:07:50,760 --> 00:07:51,760
quite opaque.

116
00:07:51,760 --> 00:07:56,360
I think even one of the challenges we sometimes face on the podcast is trying to walk that

117
00:07:56,360 --> 00:08:03,040
line between transformer architectures and retrieval augmented generation and all of

118
00:08:03,040 --> 00:08:06,920
this sort of buzzword bingo that most people are probably like, I don't know what that

119
00:08:06,920 --> 00:08:11,480
is so you've just alienated me and I don't really care about how it all works.

120
00:08:11,480 --> 00:08:13,320
I want to know what I can do with it.

121
00:08:13,320 --> 00:08:17,280
It sounds like they've bridged that gap to try and make sure that people who are not

122
00:08:17,280 --> 00:08:19,440
inside the system can still benefit.

123
00:08:19,440 --> 00:08:23,360
But I guess the other question I would have is how instructive is the document?

124
00:08:23,360 --> 00:08:29,000
Is it like, hey, you should use AI lawfully and ethically and then everyone's like, yeah,

125
00:08:29,000 --> 00:08:30,000
that makes sense.

126
00:08:30,000 --> 00:08:33,680
But then does it have instruction on how best to do that or does it not really go into that

127
00:08:33,680 --> 00:08:35,720
detail?

128
00:08:35,720 --> 00:08:43,640
There's lots of directional and kind of signposting to other documents and talking about like,

129
00:08:43,640 --> 00:08:47,720
here's a document that we've published about cyber security and here's a document about

130
00:08:47,720 --> 00:08:49,080
this that we've published.

131
00:08:49,080 --> 00:08:51,120
It is a bit instructional as well.

132
00:08:51,120 --> 00:08:58,000
Each of the core principles has a page or two going into more depth about it.

133
00:08:58,000 --> 00:09:00,680
So yeah, like I say, it's a good practical document.

134
00:09:00,680 --> 00:09:01,680
It's not short.

135
00:09:01,680 --> 00:09:04,120
It's a 70 odd page resource.

136
00:09:04,120 --> 00:09:09,520
This isn't like a five page pamphlet that you'd sit and flick through in your doctor's

137
00:09:09,520 --> 00:09:10,800
waiting room kind of thing.

138
00:09:10,800 --> 00:09:16,360
There is a little bit of detail in there that is very helpful.

139
00:09:16,360 --> 00:09:21,840
So would you advise then for SMEs, for example, thinking about, oh, crumbs, at some point

140
00:09:21,840 --> 00:09:25,400
I'm going to have to have a look at this generative AI thing and start to think about how we use

141
00:09:25,400 --> 00:09:26,400
it in our business.

142
00:09:26,400 --> 00:09:30,440
Could this be like a good first step for them to read this document?

143
00:09:30,440 --> 00:09:31,440
Very much so.

144
00:09:31,440 --> 00:09:38,720
And not just SMEs, you know, any organization, if you're a senior executive in enterprise,

145
00:09:38,720 --> 00:09:46,640
this document is as instructive for you in your role as it is for an SME entrepreneur,

146
00:09:46,640 --> 00:09:52,120
owner, managing a team of 20 to 200, whatever it may be.

147
00:09:52,120 --> 00:09:53,640
20,000 people.

148
00:09:53,640 --> 00:09:58,560
Think about how many people are in the NHS, right?

149
00:09:58,560 --> 00:10:02,120
This is just some really solid principles.

150
00:10:02,120 --> 00:10:04,560
The 10 core principles that they've come up with are very good.

151
00:10:04,560 --> 00:10:10,400
There's also some other bits in there that are really, what's the best word to describe

152
00:10:10,400 --> 00:10:12,640
it?

153
00:10:12,640 --> 00:10:15,400
Grounded in the present reality.

154
00:10:15,400 --> 00:10:20,380
So there's a section about procurement and how to procure generative AI.

155
00:10:20,380 --> 00:10:27,620
And there's a whole page on it, which is basically procuring technologies in an emerging market.

156
00:10:27,620 --> 00:10:32,240
And it talks about the risks and the limitations and how do you assess vendors when new ones

157
00:10:32,240 --> 00:10:33,720
are popping up all of the time.

158
00:10:33,720 --> 00:10:35,480
Now, this is only a page or two.

159
00:10:35,480 --> 00:10:41,200
It's not a whole textbook on it, but they're clearly recognizing that, look, this is fast

160
00:10:41,200 --> 00:10:42,200
paced.

161
00:10:42,200 --> 00:10:43,200
This is changing.

162
00:10:43,200 --> 00:10:49,560
That if you buy something today, the situation could be completely different in six months

163
00:10:49,560 --> 00:10:50,560
time.

164
00:10:50,560 --> 00:10:54,000
So you've got to make sure that you're identifying the right use cases.

165
00:10:54,000 --> 00:10:58,000
You're evaluating the technology using some really solid principles.

166
00:10:58,000 --> 00:10:59,000
Yeah.

167
00:10:59,000 --> 00:11:00,360
Overall, a great document.

168
00:11:00,360 --> 00:11:01,800
Yeah, I love that.

169
00:11:01,800 --> 00:11:03,040
I think that's some of the critical things.

170
00:11:03,040 --> 00:11:06,080
We talk about some of those things on the podcast quite often, but I think they are

171
00:11:06,080 --> 00:11:12,580
critical when it comes to getting started, figuring out the right tools.

172
00:11:12,580 --> 00:11:16,920
As we talked about a couple of times, there's so many strands to this.

173
00:11:16,920 --> 00:11:22,600
Technology is a key driver and the emerging tools and the emerging capabilities, but the

174
00:11:22,600 --> 00:11:25,840
human part of this, the change management part is just as critical.

175
00:11:25,840 --> 00:11:32,040
So yeah, nice to see a document that is a good sign poster for internal use within the

176
00:11:32,040 --> 00:11:35,360
UK government and associated infrastructure.

177
00:11:35,360 --> 00:11:39,280
But also it sounds like Martin, a document that others can use as well to try and get

178
00:11:39,280 --> 00:11:42,800
some insights into how they can move ahead with generative AI.

179
00:11:42,800 --> 00:11:43,800
Yeah.

180
00:11:43,800 --> 00:11:47,920
I'll make sure I pop a link in the show notes if anyone wants to access the document themselves.

181
00:11:47,920 --> 00:11:52,560
Well, I think that would be extremely kind of you and a good idea.

182
00:11:52,560 --> 00:11:53,560
Cool story.

183
00:11:53,560 --> 00:11:54,560
Thanks for finding that one, Martin.

184
00:11:54,560 --> 00:11:55,680
Let's move on to our next one.

185
00:11:55,680 --> 00:12:02,000
So Publicis Group, which is a pretty large global advertising communications company

186
00:12:02,000 --> 00:12:06,480
that some of our listeners might have heard of, have announced over the last week or two

187
00:12:06,480 --> 00:12:13,960
an ambitious investment of over 300 million euros in AI over the next three years, alongside

188
00:12:13,960 --> 00:12:17,760
the introduction of its new AI platform, which it calls Core AI.

189
00:12:17,760 --> 00:12:23,000
The company aims to enhance its capabilities across all of its organization with investment

190
00:12:23,000 --> 00:12:27,760
in both technology on the AI side, but also personnel development.

191
00:12:27,760 --> 00:12:33,000
Core AI is the centerpiece of Publicis Group's AI strategy, and it's designed as like an

192
00:12:33,000 --> 00:12:37,360
intelligent system that's going to incorporate things like foundational AI models, AI agents,

193
00:12:37,360 --> 00:12:41,840
there's going to be an API layer for moving information around, and also partnerships

194
00:12:41,840 --> 00:12:47,320
with third party AI services like OpenAI, Adobe, Amazon, Microsoft, people we often

195
00:12:47,320 --> 00:12:53,840
talk about here on the podcast, to basically connect trillions of data points and a database

196
00:12:53,840 --> 00:13:01,800
of 2.3 billion consumer profiles to influence how the agency groups do business, right?

197
00:13:01,800 --> 00:13:06,520
In terms of insights and strategy, media planning, creative and production, software development,

198
00:13:06,520 --> 00:13:10,220
operations, and all that good stuff.

199
00:13:10,220 --> 00:13:15,520
The market liked this because Publicis Group's stock price reached a record high off the

200
00:13:15,520 --> 00:13:17,360
back of this news.

201
00:13:17,360 --> 00:13:23,120
I think for people like you and I, Martin, you and I, you're an ex-agency and I currently

202
00:13:23,120 --> 00:13:29,640
lead an agency, it's interesting to see these big agency groups, some of them, in this case

203
00:13:29,640 --> 00:13:32,720
Publicis, going all in on AI.

204
00:13:32,720 --> 00:13:34,600
I can't say I'm surprised.

205
00:13:34,600 --> 00:13:42,720
I think it probably does take organizations or groups with this heft, right?

206
00:13:42,720 --> 00:13:47,400
This size of company to try and bake AI in in this way.

207
00:13:47,400 --> 00:13:49,000
But yeah, it's an interesting one.

208
00:13:49,000 --> 00:13:51,480
What were your thoughts on this news item?

209
00:13:51,480 --> 00:13:56,440
Not surprising when you see that they've got huge amounts of data that they're sat on.

210
00:13:56,440 --> 00:14:01,120
Digital has been at the core of the business model of these huge agencies.

211
00:14:01,120 --> 00:14:08,320
They've typically been going around acquiring businesses in the data insight and analytics

212
00:14:08,320 --> 00:14:09,680
world for a while.

213
00:14:09,680 --> 00:14:12,440
So they've got lots of top talent, typically speaking.

214
00:14:12,440 --> 00:14:19,600
I'm not sure specifically about the acquisitions of Publicis, but you see them in headlines

215
00:14:19,600 --> 00:14:23,200
about acquiring those kind of organizations regularly.

216
00:14:23,200 --> 00:14:28,520
I think this makes sense if you're going to get a competitive edge, then you've got to

217
00:14:28,520 --> 00:14:34,560
make the data that you've got available to you work for you and how else can you do that

218
00:14:34,560 --> 00:14:38,080
other than building your own AI layer.

219
00:14:38,080 --> 00:14:42,320
Makes sense that they've productized it and given it a brand name, Core AI, gives them

220
00:14:42,320 --> 00:14:47,320
something to take to the market and hang their hat on when they're going to put pictures

221
00:14:47,320 --> 00:14:51,360
and to be able to say, we've got our Core AI infrastructure.

222
00:14:51,360 --> 00:14:56,940
I'd love to see what that actually looks like.

223
00:14:56,940 --> 00:15:00,320
How is that actually manifesting itself within the team?

224
00:15:00,320 --> 00:15:09,920
Is it a front end system that people can plug into or is that more of a process and a way

225
00:15:09,920 --> 00:15:12,680
of operating?

226
00:15:12,680 --> 00:15:15,680
What does it mean?

227
00:15:15,680 --> 00:15:19,760
They talked about the API layer, so there's got to be some actual technology underpinning

228
00:15:19,760 --> 00:15:21,480
it all.

229
00:15:21,480 --> 00:15:27,760
It's not just plugging a bunch of tools into one another.

230
00:15:27,760 --> 00:15:34,440
As with all of these things, without getting your hands on it, you can't say much more

231
00:15:34,440 --> 00:15:38,400
than, oh, it's an interesting sounding press release that makes their stock price go up.

232
00:15:38,400 --> 00:15:46,120
I would welcome the opportunity to speak to some of the AI team at Publicis to talk about

233
00:15:46,120 --> 00:15:49,400
what their vision is for it and where it goes from here.

234
00:15:49,400 --> 00:15:51,800
Yeah, I think you're right.

235
00:15:51,800 --> 00:15:57,560
A lot of this, I think, is that branded AI inside play to help them differentiate themselves

236
00:15:57,560 --> 00:16:01,340
against the other big agency groups until the other big agency groups make these types

237
00:16:01,340 --> 00:16:03,760
of announcements, which are probably not far off.

238
00:16:03,760 --> 00:16:07,680
I think you're right in terms of how data centric a lot of these groups have become

239
00:16:07,680 --> 00:16:08,680
over the years.

240
00:16:08,680 --> 00:16:15,680
It's almost like their mini but slightly diversified Facebook approach.

241
00:16:15,680 --> 00:16:16,800
Facebook's got loads of data on us.

242
00:16:16,800 --> 00:16:21,240
They're using AI to help surface content and show ads to the right people and even

243
00:16:21,240 --> 00:16:23,480
to generate ads now.

244
00:16:23,480 --> 00:16:29,480
You could definitely imagine how Publicis Group could use similar tools internally for

245
00:16:29,480 --> 00:16:33,820
audience creation and segmentation for their clients, predictions in terms of consumer

246
00:16:33,820 --> 00:16:39,240
purchase habits and then building strategies, content creation, programmatic advertising

247
00:16:39,240 --> 00:16:43,720
off the back of that.

248
00:16:43,720 --> 00:16:47,440
I don't even know how much their clients are going to get to see of how that machinery

249
00:16:47,440 --> 00:16:48,440
works if I'm honest.

250
00:16:48,440 --> 00:16:51,840
I think it'll be more, these are the outcomes that we can get.

251
00:16:51,840 --> 00:16:54,720
These are the outputs that we'll produce.

252
00:16:54,720 --> 00:16:59,040
There's AI magic in here and AI is buzz and cool and we're buzz and cool and you should

253
00:16:59,040 --> 00:17:00,040
work with us.

254
00:17:00,040 --> 00:17:03,600
I suspect this is where a lot of that will come down to.

255
00:17:03,600 --> 00:17:04,600
There are some tools.

256
00:17:04,600 --> 00:17:07,600
It's nice that they're creating their own infrastructure and tools to use but there

257
00:17:07,600 --> 00:17:10,920
are some tools that hopefully lots of people are going to be able to use like Google's

258
00:17:10,920 --> 00:17:14,320
new text to video model, Martin.

259
00:17:14,320 --> 00:17:19,440
Yeah, nice to see Google Research publishing new papers.

260
00:17:19,440 --> 00:17:27,520
They've just developed a new model called Lumiere which is a space-time text to video

261
00:17:27,520 --> 00:17:34,440
diffusion model capable of transforming text and images into realistic AI generated videos.

262
00:17:34,440 --> 00:17:41,280
Now I am not going to pretend that I know what space-time text to video diffusion model

263
00:17:41,280 --> 00:17:46,760
actually means but what I can tell you is that what makes this model slightly different

264
00:17:46,760 --> 00:17:52,840
is it stands out for its ability to generate full frame rate, low resolution videos in

265
00:17:52,840 --> 00:18:01,000
one pass using what they're calling the space-time unit architecture which processes multiple

266
00:18:01,000 --> 00:18:03,880
space-time scales.

267
00:18:03,880 --> 00:18:12,280
I read this story and felt overwhelmed with technical input.

268
00:18:12,280 --> 00:18:15,140
I was like, I need someone to explain this like I'm five.

269
00:18:15,140 --> 00:18:20,840
So I asked Perplexity to explain it like I'm five and it did a good job but it basically

270
00:18:20,840 --> 00:18:29,080
did just say this is a computer system that can make videos and it looks at an image and

271
00:18:29,080 --> 00:18:30,080
it makes an image.

272
00:18:30,080 --> 00:18:31,080
There you go.

273
00:18:31,080 --> 00:18:32,080
Right, right.

274
00:18:32,080 --> 00:18:33,080
Okay.

275
00:18:33,080 --> 00:18:38,040
Maybe you've dumbed it down slightly too hard for me there.

276
00:18:38,040 --> 00:18:42,000
But yeah, this is an interesting development.

277
00:18:42,000 --> 00:18:46,160
Having seen the demo video, again demo videos, right, always got to be wary particularly

278
00:18:46,160 --> 00:18:50,920
with Google's recent track record of demo videos and people being scandalized but the

279
00:18:50,920 --> 00:18:52,640
demo video looks cool.

280
00:18:52,640 --> 00:18:57,120
The videos look high quality.

281
00:18:57,120 --> 00:19:02,600
They've trained it on a data set of 30 million videos and text captions.

282
00:19:02,600 --> 00:19:07,640
They've not disclosed any of the training data set but let's be honest, they've got

283
00:19:07,640 --> 00:19:11,340
enough data that they're sat on.

284
00:19:11,340 --> 00:19:17,960
It can do text to video generation, converts still images into videos, generates videos

285
00:19:17,960 --> 00:19:22,640
in specific styles using a reference image and it does video inpainting.

286
00:19:22,640 --> 00:19:28,360
So its capabilities are quite broad, certainly much broader than we see from some of the

287
00:19:28,360 --> 00:19:32,040
other text to image generators on the market.

288
00:19:32,040 --> 00:19:37,760
It produces five second videos at the moment and it's not available for us to use.

289
00:19:37,760 --> 00:19:43,640
This is very much just a research paper but I expect that we'll see some things on the

290
00:19:43,640 --> 00:19:53,080
market in the near future based on how long did it take for the imaging to come to market.

291
00:19:53,080 --> 00:19:56,040
It was about the best part of a year wasn't it actually?

292
00:19:56,040 --> 00:20:00,880
Yeah, I think my gut on this stuff is that sometimes we see the products accessible in

293
00:20:00,880 --> 00:20:05,600
some sort of demo form in about three months but in terms of it being baked into something

294
00:20:05,600 --> 00:20:11,480
that you could use on a sort of more regular basis, yeah six months maybe 12.

295
00:20:11,480 --> 00:20:12,480
So hopefully it will come soon.

296
00:20:12,480 --> 00:20:18,240
I mean watching the demo videos and trying to understand this space-time which just sounds

297
00:20:18,240 --> 00:20:19,400
really cool anyway.

298
00:20:19,400 --> 00:20:22,400
Like let's give them props for that.

299
00:20:22,400 --> 00:20:27,780
I feel like I'm in Star Wars or Star Trek or something.

300
00:20:27,780 --> 00:20:31,820
Is this ability to create more realistic motion?

301
00:20:31,820 --> 00:20:37,280
Because one of the issues when you're trying to create motion in tools like Runway is people's

302
00:20:37,280 --> 00:20:44,360
heads morph weirdly, even items in the videos morph weirdly.

303
00:20:44,360 --> 00:20:50,640
So it's kind of hard to get a moving image to video generation where the original image

304
00:20:50,640 --> 00:20:54,040
doesn't get mangled in some way during the process so that you look at it and you're

305
00:20:54,040 --> 00:20:56,840
like that's clearly AI generated.

306
00:20:56,840 --> 00:21:00,480
Even the best examples for me are super obviously AI generated.

307
00:21:00,480 --> 00:21:06,320
I think what reading between the lines here the nature of how this model works should

308
00:21:06,320 --> 00:21:16,360
ensure that frame to frame the image is animating or sort of moving in a way that's more realistic.

309
00:21:16,360 --> 00:21:20,820
So if someone turns their head it should look like they're normally turning their head.

310
00:21:20,820 --> 00:21:25,400
If their hair blows in the breeze it should look more natural.

311
00:21:25,400 --> 00:21:29,440
If a bow is moving left to right it should look more natural and I think that's what

312
00:21:29,440 --> 00:21:30,440
they're claiming.

313
00:21:30,440 --> 00:21:34,720
So I think you said until we can actually get our hands on it and play with it because

314
00:21:34,720 --> 00:21:40,200
the gap there will be a gap between the demos and the real world use because as we talk

315
00:21:40,200 --> 00:21:45,420
about a lot and this might be user error more than anything but you and I struggle to get

316
00:21:45,420 --> 00:21:49,360
anything that looks particularly good out of Runway if we're honest.

317
00:21:49,360 --> 00:21:54,380
I think you have to spend ages refining your prompt and thinking carefully about your input

318
00:21:54,380 --> 00:21:57,240
image to get stuff that looks really good.

319
00:21:57,240 --> 00:22:02,120
I would love this to be a step forward in usability for your average user so you can

320
00:22:02,120 --> 00:22:06,960
just chuck an image in put in a prompt to get a half decent result without all of the

321
00:22:06,960 --> 00:22:12,980
garbledness that seems to plague most video generation tools at the moment.

322
00:22:12,980 --> 00:22:13,980
That is the dream.

323
00:22:13,980 --> 00:22:19,720
While we're on video generation tools and we've mentioned Runway who we think are awesome

324
00:22:19,720 --> 00:22:24,200
by the way and doing great work in this area they are not resting on their laurels either

325
00:22:24,200 --> 00:22:30,600
because they have upgraded their AI video generation capabilities with the new multi-motion

326
00:22:30,600 --> 00:22:34,920
brush in their flagship Gen 2 video generation model.

327
00:22:34,920 --> 00:22:40,400
So many of you might know from previous podcast episodes when we discussed the introduction

328
00:22:40,400 --> 00:22:44,880
of the original motion brush which was interesting because rather than just uploading an image

329
00:22:44,880 --> 00:22:50,120
as a prompt and then Gen 2 tries to animate the whole image you could use the motion brush

330
00:22:50,120 --> 00:22:55,000
to say to just paint a certain bit of the image and just encourage the model to only

331
00:22:55,000 --> 00:22:56,000
animate that.

332
00:22:56,000 --> 00:23:01,560
So an example would be a boat on a river but you don't want but there's loads of other

333
00:23:01,560 --> 00:23:04,880
things that are in the shot but you don't want those to animate so you just paint the

334
00:23:04,880 --> 00:23:08,040
boat and then ideally the boat sails along on the river.

335
00:23:08,040 --> 00:23:13,160
Well now with the multi-brush multi-motion brush feature instead of just painting one

336
00:23:13,160 --> 00:23:18,920
thing you can now paint multiple things that will animate in different ways.

337
00:23:18,920 --> 00:23:23,880
So you might want one boat going one way but you might want another boat going in the opposite

338
00:23:23,880 --> 00:23:29,440
direction which in theory now you can do because you can paint the image in essence with different

339
00:23:29,440 --> 00:23:33,720
brushes in different colours and have them do different things.

340
00:23:33,720 --> 00:23:38,840
So this is kind of cool because it's part of the ongoing evolution of how these video

341
00:23:38,840 --> 00:23:44,680
tools are maturing giving users ever greater control about which parts of an image assuming

342
00:23:44,680 --> 00:23:48,680
they're using an image in their prompt that they want it to animate.

343
00:23:48,680 --> 00:23:53,560
I haven't actually had a play with the multi-motion brush multi-brush motion crumbs it's a bit

344
00:23:53,560 --> 00:23:54,560
of a mouthful.

345
00:23:54,560 --> 00:23:56,400
I don't know if you've had a play with this yet.

346
00:23:56,400 --> 00:24:00,080
Mark I've seen some of the videos that look pretty cool but I haven't actually tried it

347
00:24:00,080 --> 00:24:03,000
out and broken it like I probably will.

348
00:24:03,000 --> 00:24:08,480
Not with the multi-motion I tried the motion brush and just failed.

349
00:24:08,480 --> 00:24:16,320
I think with these video generation models there's a...

350
00:24:16,320 --> 00:24:20,440
Okay as with any AI you've got to understand the limitations and where it's currently at

351
00:24:20,440 --> 00:24:29,160
and I think I still not found where that kind of sweet spot is when I'm using these tools.

352
00:24:29,160 --> 00:24:36,280
I understand the limits of what GPT-3 can do and I understand how to get good value

353
00:24:36,280 --> 00:24:40,400
and good outputs from GPT-4.

354
00:24:40,400 --> 00:24:44,680
Darley and stable diffusion and mid-journey I'm getting good results from those and I

355
00:24:44,680 --> 00:24:50,200
know where I can push the boundaries and where to row back.

356
00:24:50,200 --> 00:24:57,880
But video I can't get there and I try and I try different prompting techniques and it

357
00:24:57,880 --> 00:24:59,320
does take time.

358
00:24:59,320 --> 00:25:04,360
So I think there's definitely an element of user error about this but I also think the

359
00:25:04,360 --> 00:25:11,760
models, the video models just still are a little bit...

360
00:25:11,760 --> 00:25:19,600
And that's probably why we're not seeing lots and lots of examples coming out of people

361
00:25:19,600 --> 00:25:20,600
creating great stuff.

362
00:25:20,600 --> 00:25:25,800
And actually when I see a video come out and bearing in mind we're both in these communities

363
00:25:25,800 --> 00:25:32,200
online when you do see some of these videos come out they're kind of noticeable because

364
00:25:32,200 --> 00:25:34,840
they're an exception.

365
00:25:34,840 --> 00:25:42,440
I think a lot of people are having similar struggles in finding where these models work

366
00:25:42,440 --> 00:25:47,120
well and where they fall apart because I'm really good at finding where they fall apart.

367
00:25:47,120 --> 00:25:53,400
Yeah I think you've nailed it there because the best examples I see are people creating

368
00:25:53,400 --> 00:25:59,120
sort of trailers or like very short movies that are like 60 seconds long that are fast

369
00:25:59,120 --> 00:26:05,120
cutting because these tools can only really produce two, three, five seconds of video

370
00:26:05,120 --> 00:26:09,400
that has like it starts to lose its coherence the longer it goes.

371
00:26:09,400 --> 00:26:10,800
And so they're cutting quite quickly.

372
00:26:10,800 --> 00:26:17,920
They're trying to maintain consistent characters in between different shots if you like.

373
00:26:17,920 --> 00:26:20,720
And I suspect that in most cases they're using image prompting.

374
00:26:20,720 --> 00:26:26,520
So the image is in the prompt to help really direct the system on how which, what motion

375
00:26:26,520 --> 00:26:31,120
to create what the video output video should look like.

376
00:26:31,120 --> 00:26:34,600
For me this is still in the hobbyist have some fun with it phase.

377
00:26:34,600 --> 00:26:39,440
I don't think it's got the necessary production level power to be of real use to marketers.

378
00:26:39,440 --> 00:26:43,240
I'm going to stick my neck out and say that.

379
00:26:43,240 --> 00:26:48,720
My use case that I expect to come first but again I'm frustrated because I can't get

380
00:26:48,720 --> 00:26:54,320
a decent output reasonably quickly is just adding some motion to things like social media

381
00:26:54,320 --> 00:26:58,680
static images that would have been static but it would be really great to just animate

382
00:26:58,680 --> 00:27:01,440
a few bits of them.

383
00:27:01,440 --> 00:27:05,840
To be honest the go-to for me there is still probably the little animations that Canva

384
00:27:05,840 --> 00:27:08,540
can do which are very templated.

385
00:27:08,540 --> 00:27:12,440
They look like sort of standard PowerPoint slide type animations.

386
00:27:12,440 --> 00:27:13,840
They're not particularly interesting.

387
00:27:13,840 --> 00:27:19,360
I think when you can take a branded image introduce a bit of motion in runway and then

388
00:27:19,360 --> 00:27:25,740
run that on social there'll be a time period where that's interesting and you'll get attention

389
00:27:25,740 --> 00:27:29,640
online because you're doing it and then it will become passe and everybody's images will

390
00:27:29,640 --> 00:27:32,160
do it and then no one will pay attention anymore.

391
00:27:32,160 --> 00:27:35,840
So that's the sort of that's I think going to be the first use case that as a marketer

392
00:27:35,840 --> 00:27:40,160
unlocks it for me but as I say it's far too much effort to get anything half decent for

393
00:27:40,160 --> 00:27:43,120
me at least out of it for even that simple use case.

394
00:27:43,120 --> 00:27:46,280
So until we cross that barrier I'm playing with it.

395
00:27:46,280 --> 00:27:49,640
I'm not thinking about how I can use it for work if I'm honest.

396
00:27:49,640 --> 00:27:51,880
Yeah exactly the same.

397
00:27:51,880 --> 00:27:56,920
Right let's move on from video and let's talk a little bit about OpenAI's announcements

398
00:27:56,920 --> 00:27:57,920
this week.

399
00:27:57,920 --> 00:28:03,480
Martin tell us a bit more about these enhancements and cost reductions that were shared on the

400
00:28:03,480 --> 00:28:04,760
Twittersphere.

401
00:28:04,760 --> 00:28:08,240
Small changes but significant ones nonetheless.

402
00:28:08,240 --> 00:28:14,680
They've announced some updates on their text embeddings models as well as some improvements

403
00:28:14,680 --> 00:28:21,760
and optimizations to GPT-4 Turbo and GPT-3.5 Turbo.

404
00:28:21,760 --> 00:28:28,520
Since GPT-4 Turbo was launched one of the criticisms especially from developers was

405
00:28:28,520 --> 00:28:31,020
that it had become kind of lazy.

406
00:28:31,020 --> 00:28:37,040
If you asked it to write some code it would not give you the full coding.

407
00:28:37,040 --> 00:28:38,680
Truncate the code.

408
00:28:38,680 --> 00:28:43,960
It would tell you to add in bits here and there and put lots of comments in it saying

409
00:28:43,960 --> 00:28:47,360
and this is where you need to integrate this with that.

410
00:28:47,360 --> 00:28:49,680
And sometimes you just wanted it to give me the full code.

411
00:28:49,680 --> 00:28:53,680
I've given you 40 lines of code.

412
00:28:53,680 --> 00:28:59,700
I wanted you to fix the bug and give me the full 40 lines of code back fixed.

413
00:28:59,700 --> 00:29:04,440
But instead what it does is it gives you like 10 lines and says copy and paste this here.

414
00:29:04,440 --> 00:29:06,520
Well it sounds like they've fixed that.

415
00:29:06,520 --> 00:29:13,680
They've made it less lazy and giving it a more complete output.

416
00:29:13,680 --> 00:29:21,560
They've also improved some issues that were reported with non-English UTF-8 generation.

417
00:29:21,560 --> 00:29:24,160
So this is some of the characters that it would generate.

418
00:29:24,160 --> 00:29:28,920
It wasn't very good and it would be a bit buggy.

419
00:29:28,920 --> 00:29:37,200
So GPT-3.5 Turbo they've made some improvements here on the cost front.

420
00:29:37,200 --> 00:29:47,080
So now the model offers 50% lower input prices and 25% reduced output prices making the model

421
00:29:47,080 --> 00:29:49,640
more affordable.

422
00:29:49,640 --> 00:29:54,600
And I think they've made some improvements to the model in terms of accuracy and format

423
00:29:54,600 --> 00:29:55,840
response capabilities.

424
00:29:55,840 --> 00:30:01,860
But the big headline here is it's substantially cheaper to use than it was previously.

425
00:30:01,860 --> 00:30:02,860
And that matters.

426
00:30:02,860 --> 00:30:11,520
I've been doing a project recently where we've been working with a client to help improve

427
00:30:11,520 --> 00:30:18,600
their analysis and insight extraction from lots of qualitative data.

428
00:30:18,600 --> 00:30:24,800
Spreadsheets with 70 rows of text inputs.

429
00:30:24,800 --> 00:30:32,160
Probably 30 columns of data formatted badly, completely inconsistent in terms of the responses

430
00:30:32,160 --> 00:30:35,080
you get from people.

431
00:30:35,080 --> 00:30:40,700
And using the GPT for Sheets plugin, which is excellent by the way, if anybody wants

432
00:30:40,700 --> 00:30:48,040
to integrate the OpenAI models with Google Sheets, I highly recommend that as a plugin.

433
00:30:48,040 --> 00:30:50,280
But also integrating the Claude for Sheets plugin.

434
00:30:50,280 --> 00:30:57,080
I've been playing around trying to get insight extraction and it gets expensive very quickly.

435
00:30:57,080 --> 00:31:04,640
Especially when every time you load the sheet, so every time you open it, unless you can

436
00:31:04,640 --> 00:31:08,800
turn on caching, so it stores the data for six hours, but if you don't have that turned

437
00:31:08,800 --> 00:31:13,800
on, it refreshes every cell and runs the prompt again.

438
00:31:13,800 --> 00:31:21,240
And I found that when I had GPT-4, not GPT-4 Turbo turned on, and GPT-4 Turbo is considerably

439
00:31:21,240 --> 00:31:32,520
cheaper than GPT-4, just refreshing the spreadsheet once ended up costing me a couple of quid.

440
00:31:32,520 --> 00:31:34,620
So just to run it once.

441
00:31:34,620 --> 00:31:38,560
So that is quite expensive and not something that you want to do regularly.

442
00:31:38,560 --> 00:31:46,880
But yeah, GPT-3, GPT-4, if we can get the prices down, that's going to be better for

443
00:31:46,880 --> 00:31:48,480
everyone.

444
00:31:48,480 --> 00:31:56,080
Just on the insights extraction piece actually, what was fascinating to me was having columns

445
00:31:56,080 --> 00:32:08,400
with the analysis of each row from GPT-3.5, GPT-4 and Claude 2.1.

446
00:32:08,400 --> 00:32:17,200
The difference in quality and detail with the exact verbatim same prompts was really,

447
00:32:17,200 --> 00:32:18,600
really insightful.

448
00:32:18,600 --> 00:32:25,920
GPT-4, and this should come as no surprise, was far and away the more detailed, the more

449
00:32:25,920 --> 00:32:28,280
insightful, the more nuanced.

450
00:32:28,280 --> 00:32:38,600
Claude 2.1, I would say was very, very, very similar to GPT-3.5.

451
00:32:38,600 --> 00:32:44,280
If anything, maybe marginally fractionally worse, but that could be down to the style

452
00:32:44,280 --> 00:32:45,280
of the prompt.

453
00:32:45,280 --> 00:32:51,480
I mean, they were basically on a par, but the difference between 3.5 and 4 was night

454
00:32:51,480 --> 00:32:52,680
and day.

455
00:32:52,680 --> 00:32:58,600
The frustration from a practical perspective, if anybody is going to implement GPT for Sheets,

456
00:32:58,600 --> 00:33:06,560
is be aware that Google Sheets has a 30 second timeout if you're making an external API call.

457
00:33:06,560 --> 00:33:12,840
So without getting too technical on it, when you run that query and it sends that query

458
00:33:12,840 --> 00:33:19,640
and you're prompt off to open AI, if it doesn't get a response back within 30 seconds, you

459
00:33:19,640 --> 00:33:21,640
see an error message.

460
00:33:21,640 --> 00:33:29,280
And the problem with GPT-4 and GPT-4 Turbo is the latency on these models is quite slow.

461
00:33:29,280 --> 00:33:36,680
So if you've got quite a long prompt and you're expecting quite a long response, it can be

462
00:33:36,680 --> 00:33:39,920
more than 30 seconds, therefore you'll get an error message.

463
00:33:39,920 --> 00:33:43,920
And this can, let me tell you, be infuriating.

464
00:33:43,920 --> 00:33:46,080
Yeah, for those...

465
00:33:46,080 --> 00:33:50,440
We've talked about this quite a few months ago, so if you're new to the podcast, you

466
00:33:50,440 --> 00:33:55,000
might not have heard this story, but in the original days of the podcast, one of the key

467
00:33:55,000 --> 00:33:59,980
drivers for Martin and I was we want to use AI in as many areas of podcast production

468
00:33:59,980 --> 00:34:03,840
as we can, because that's kind of half the point of doing the podcast, is to see what

469
00:34:03,840 --> 00:34:05,420
can AI help with, right?

470
00:34:05,420 --> 00:34:09,900
Having two humans having a conversation should be one of the most human driven things that

471
00:34:09,900 --> 00:34:13,760
can exist, but we found ways to infuse AI in a number of areas.

472
00:34:13,760 --> 00:34:19,160
And one of the things that we tried was using Google Sheets, using a plugin like this and

473
00:34:19,160 --> 00:34:25,560
having GPT-4 write the script for the podcast based on the stories that we copy paste into

474
00:34:25,560 --> 00:34:26,560
it.

475
00:34:26,560 --> 00:34:31,080
So in cell 1A, you copy paste your story and then this is a simplification, but in cell

476
00:34:31,080 --> 00:34:39,920
1B, it sends that story to GPT-4 with a set of instructions and GPT-4 writes a script

477
00:34:39,920 --> 00:34:45,520
off the back of it and then it populates that cell, which was great, except three quarters

478
00:34:45,520 --> 00:34:49,040
of the time it would time out, which basically made it unusable, right?

479
00:34:49,040 --> 00:34:51,680
Because it was supposed to be something that would save us time, but you spent so many

480
00:34:51,680 --> 00:34:56,080
times rerunning, rerunning, rerunning the prompt until you actually got an output, whilst

481
00:34:56,080 --> 00:35:00,300
burning API costs the whole time.

482
00:35:00,300 --> 00:35:04,560
So that 30 second window is a pain and it causes a lot of issues, but if you're listening

483
00:35:04,560 --> 00:35:09,640
to this, I think Martin's use case there is really awesome and I've had a number of people

484
00:35:09,640 --> 00:35:14,360
when I've talked about this out and about really thinking of interesting ways to use

485
00:35:14,360 --> 00:35:19,520
spreadsheet data to automatically drive GPT-4 outputs.

486
00:35:19,520 --> 00:35:25,560
So another great example would be you want to do some sort of cold email outbound prospecting

487
00:35:25,560 --> 00:35:30,960
process, but you want your emails to be somewhat customized, while now if you have column A

488
00:35:30,960 --> 00:35:36,200
has maybe the person's name, column B has some information about their company and column

489
00:35:36,200 --> 00:35:42,800
C has the particular product that you think they'll be interested in, you could now create

490
00:35:42,800 --> 00:35:49,240
customized prospecting emails that are more than just the smart content that we use in

491
00:35:49,240 --> 00:35:51,120
a lot of the automation systems today.

492
00:35:51,120 --> 00:35:54,960
You know, hi first name, how's it going over at company, right?

493
00:35:54,960 --> 00:35:57,480
They're going to be better than that.

494
00:35:57,480 --> 00:36:03,280
So I think we're back to these are the tools, this is the practical thing they can do, right?

495
00:36:03,280 --> 00:36:09,880
They can run prompts and generate information from GPT-4 at scale.

496
00:36:09,880 --> 00:36:14,520
How are you going to structure a spreadsheet to drive value in your marketing and sales

497
00:36:14,520 --> 00:36:20,480
organization using it, whilst taking on board Martin's really helpful comments about the

498
00:36:20,480 --> 00:36:27,400
levels in quality difference between GPT-4, 3.5 and Claude, the cost differences, you

499
00:36:27,400 --> 00:36:32,720
know, you've got to get that cost benefit ratio right and that darn 30 second timeout

500
00:36:32,720 --> 00:36:37,280
that's going to trip you up if you try and do something too lengthy or complicated that

501
00:36:37,280 --> 00:36:40,600
GPT-4 just struggles to produce the information in the right time.

502
00:36:40,600 --> 00:36:43,960
Cool, thanks for sharing that with us Martin.

503
00:36:43,960 --> 00:36:47,000
Can I ask you a quick question actually?

504
00:36:47,000 --> 00:36:50,000
Who's using GPT-3.5?

505
00:36:50,000 --> 00:36:53,720
So reducing the cost for it, so they obviously think some people are going to benefit from

506
00:36:53,720 --> 00:36:54,720
that.

507
00:36:54,720 --> 00:36:55,840
Like what's the use case for that?

508
00:36:55,840 --> 00:36:58,360
Because we've just talked about how much better GPT-4 is.

509
00:36:58,360 --> 00:37:00,840
So who's using GPT-3.5 in these things?

510
00:37:00,840 --> 00:37:08,360
Yeah, so GPT-3.5 I think still does an excellent job, particularly with a well crafted prompt.

511
00:37:08,360 --> 00:37:15,080
In fact, in the project I was just describing, the next iteration for me is to rather than

512
00:37:15,080 --> 00:37:21,040
having zero shot prompting for GPT-3, what I'm going to do is take the output that I

513
00:37:21,040 --> 00:37:30,960
got from GPT-4 and give it that in the prompt for GPT-3.5 to kind of give it an example

514
00:37:30,960 --> 00:37:33,760
of the level of detail and information that I want.

515
00:37:33,760 --> 00:37:36,720
So I'm going to give it a few shot prompt.

516
00:37:36,720 --> 00:37:43,240
And I think that's where GPT-3.5 will come into its own because it's really good with

517
00:37:43,240 --> 00:37:47,280
few shot prompting, but it means your input costs are higher because you've got to do

518
00:37:47,280 --> 00:37:48,280
much longer.

519
00:37:48,280 --> 00:37:52,600
You've got to do much longer prompting.

520
00:37:52,600 --> 00:37:57,640
But when it sees that pattern recognition, it does a very good job.

521
00:37:57,640 --> 00:38:05,480
Also for a lot of applications, GPT-3.5 again with few shot prompting is really good, you

522
00:38:05,480 --> 00:38:11,020
know, writing social media posts can be very good if you've given it some examples.

523
00:38:11,020 --> 00:38:15,480
Its classification is really good for that.

524
00:38:15,480 --> 00:38:19,280
Doing sentiment analysis, really good for that.

525
00:38:19,280 --> 00:38:25,520
So I think there is a lot of use cases, but I think for most people as an assistant, you

526
00:38:25,520 --> 00:38:29,680
know, that kind of back and forth dialogue, help me solve this problem.

527
00:38:29,680 --> 00:38:31,320
GPT-4 is just much better.

528
00:38:31,320 --> 00:38:39,600
But where you have a particularly a use case like we've got the podcast, we could, we don't,

529
00:38:39,600 --> 00:38:47,960
but we could get the RSS feed from the podcast, put it into spreadsheet, taking episode title,

530
00:38:47,960 --> 00:38:55,400
show notes, publish date, etc. and just have a series of columns giving us tweets, giving

531
00:38:55,400 --> 00:39:01,080
us LinkedIn posts, giving us email subject lines in our style and our tone of voice because

532
00:39:01,080 --> 00:39:04,720
we've done that in the few shot prompt where we give the prompt examples of ones that we

533
00:39:04,720 --> 00:39:07,240
like and then it gives us the output.

534
00:39:07,240 --> 00:39:10,560
So I think that's where it can be applied quite effectively.

535
00:39:10,560 --> 00:39:14,520
That's quite cool because then we can use a tool like Zapier to just automatically schedule

536
00:39:14,520 --> 00:39:17,240
those posts.

537
00:39:17,240 --> 00:39:20,200
Ideally with a human in the loop, so whatever tool we're using, we can go and check and

538
00:39:20,200 --> 00:39:24,960
make sure they're not absolute nonsense and then they're ready for posting.

539
00:39:24,960 --> 00:39:26,360
Because nobody would ever do that, would they Paul?

540
00:39:26,360 --> 00:39:30,840
I definitely haven't done that in the past on this very podcast and shared posts on LinkedIn,

541
00:39:30,840 --> 00:39:35,160
fully automated and gone, ah, I didn't mean to do that.

542
00:39:35,160 --> 00:39:41,120
Yeah, when I had to forward that one to you and I was like, it's Friday, I'm quite tired

543
00:39:41,120 --> 00:39:43,920
Martin, but this makes no sense, does it?

544
00:39:43,920 --> 00:39:50,120
No, sorry, that was an AI generated one that slipped through.

545
00:39:50,120 --> 00:39:52,240
It happens to the best of us.

546
00:39:52,240 --> 00:39:54,520
Right, let's move on to our next story then Martin.

547
00:39:54,520 --> 00:39:59,360
We're going to talk 11 Labs briefly because they just got a ton of funding.

548
00:39:59,360 --> 00:40:04,680
So for those of you that are new to the podcast, 11 Labs is one of the leaders in voice technology,

549
00:40:04,680 --> 00:40:08,000
so they're the ones that have been creating synthetic voices where you write a script

550
00:40:08,000 --> 00:40:11,360
and then it reads it out for you and sounds very, very human.

551
00:40:11,360 --> 00:40:16,520
They've been piloting and driving voice cloning in terms of cloning your own voice and writing

552
00:40:16,520 --> 00:40:20,920
a script that then speaks in your own voice and also trying to introduce other clever

553
00:40:20,920 --> 00:40:28,960
aspects to voice synthesis around maintaining emotional style, translating between languages.

554
00:40:28,960 --> 00:40:33,240
So if you speak in English, it will then translate it and have you speak it in Spanish in your

555
00:40:33,240 --> 00:40:35,400
voice and all that good stuff.

556
00:40:35,400 --> 00:40:37,400
And this week they got $80 million.

557
00:40:37,400 --> 00:40:41,320
That felt a bit Dr. Evil there.

558
00:40:41,320 --> 00:40:49,760
$80 million in Series B funding led by notable investors like Andreessen Horowitz, Sequoia

559
00:40:49,760 --> 00:40:56,680
Capital and all these firms that are very much Silicon Valley tech investment.

560
00:40:56,680 --> 00:41:01,800
But I think this just goes to show how people see this as being big business.

561
00:41:01,800 --> 00:41:05,720
I did see some sort of, will I call it banter?

562
00:41:05,720 --> 00:41:08,040
Yeah, let's go with banter.

563
00:41:08,040 --> 00:41:13,840
But basically people questioning the value, because I think it puts their valuation at

564
00:41:13,840 --> 00:41:20,800
like half a billion or something, and people questioning how defensible their moat is when

565
00:41:20,800 --> 00:41:24,680
Meta is releasing open source models supposedly.

566
00:41:24,680 --> 00:41:27,360
We've seen some demos of what their models can do.

567
00:41:27,360 --> 00:41:29,920
No one can access them yet as far as I know.

568
00:41:29,920 --> 00:41:32,560
They've been very open sourcey so far, have Meta.

569
00:41:32,560 --> 00:41:37,760
So is there a open source voice synthesis model coming from Meta that's basically as

570
00:41:37,760 --> 00:41:39,360
good as Eleven Labs is?

571
00:41:39,360 --> 00:41:44,800
We've got PlayHT, we've got voice synthesis starting to be built into tools in a sort

572
00:41:44,800 --> 00:41:47,200
of rudimentary way like Descript.

573
00:41:47,200 --> 00:41:48,200
So I don't know.

574
00:41:48,200 --> 00:41:52,680
I think that's a fair criticism in terms of just how defensible Eleven Labs' position

575
00:41:52,680 --> 00:41:53,680
is.

576
00:41:53,680 --> 00:41:54,680
We love them, we should say as well.

577
00:41:54,680 --> 00:41:55,680
We think their technology is cool.

578
00:41:55,680 --> 00:41:56,680
We play with it all the time.

579
00:41:56,680 --> 00:41:58,360
We're always looking for cool use cases.

580
00:41:58,360 --> 00:42:03,840
But a lot of money going into a lot of AI apps that might not have a defensible market

581
00:42:03,840 --> 00:42:06,240
position, which is interesting.

582
00:42:06,240 --> 00:42:14,720
When you consider that you can do AI voice generation through Google's cloud AI suite,

583
00:42:14,720 --> 00:42:20,520
Amazon has got a system called Poly, which is just AI voice generation, PlayHT you've

584
00:42:20,520 --> 00:42:21,760
mentioned.

585
00:42:21,760 --> 00:42:24,480
You've got Lovo with their Geni application.

586
00:42:24,480 --> 00:42:27,320
You've got Eleven Labs.

587
00:42:27,320 --> 00:42:28,920
The competition is there.

588
00:42:28,920 --> 00:42:33,480
Plus, as you say, Meta are working on this as well and they're very open-sourcy in the

589
00:42:33,480 --> 00:42:34,480
way that they do it.

590
00:42:34,480 --> 00:42:40,300
So this is going to be a very competitive landscape and it is going to be baked into

591
00:42:40,300 --> 00:42:44,520
tools left, right and center.

592
00:42:44,520 --> 00:42:47,800
Half a billion does seem high.

593
00:42:47,800 --> 00:42:49,800
There's a lot of hype at the moment.

594
00:42:49,800 --> 00:42:56,000
Can they keep the momentum going with new product releases?

595
00:42:56,000 --> 00:43:03,360
They've just launched the dubbing studio, which is a voice library as well.

596
00:43:03,360 --> 00:43:06,560
They've got quite a big voice library.

597
00:43:06,560 --> 00:43:14,760
Just on that dubbing studio that offers enhanced content production capabilities where you

598
00:43:14,760 --> 00:43:18,760
can dub entire videos.

599
00:43:18,760 --> 00:43:24,400
So you can just drop in a whole clip and it will overdub a whole piece.

600
00:43:24,400 --> 00:43:32,920
They've got a mobile reader app, which facilitates the instant conversion of text and URLs into

601
00:43:32,920 --> 00:43:35,320
audio.

602
00:43:35,320 --> 00:43:42,400
They've got a very good API, which even I've managed to build things that integrate with

603
00:43:42,400 --> 00:43:43,400
it.

604
00:43:43,400 --> 00:43:47,460
I'm not a coder, as I've said many times in the past.

605
00:43:47,460 --> 00:43:52,520
So if they can keep innovating and pushing the boundaries and I think what sets them

606
00:43:52,520 --> 00:43:57,640
apart at the moment, and I've actually written a blog post about this, is the quality of

607
00:43:57,640 --> 00:44:01,360
their voices is just better than all of the others.

608
00:44:01,360 --> 00:44:13,280
The nearest I can find in terms of emotional range and having some depth to the voice is

609
00:44:13,280 --> 00:44:17,520
Lovvo and the quality of the outputs you get from Jenny.

610
00:44:17,520 --> 00:44:24,320
But 11 Labs, I think is just the best in terms of quality of voice at the moment.

611
00:44:24,320 --> 00:44:25,320
Yeah.

612
00:44:25,320 --> 00:44:28,640
I mean, and they're buzzing because I think they've only got, you know, they're less than

613
00:44:28,640 --> 00:44:32,960
a hundred employees and this big evaluation.

614
00:44:32,960 --> 00:44:34,840
So I do think they're doing exciting things.

615
00:44:34,840 --> 00:44:41,320
I should say I was summarizing some back and forth on Twitter and LinkedIn questioning

616
00:44:41,320 --> 00:44:42,320
this.

617
00:44:42,320 --> 00:44:46,880
I'm not saying that's like my own personal opinion, although I think it's an interesting

618
00:44:46,880 --> 00:44:51,160
potential marker of the amount of hype inflation that's happening in the market.

619
00:44:51,160 --> 00:44:52,800
But I think their tools are really cool.

620
00:44:52,800 --> 00:44:56,000
I completely agree with that.

621
00:44:56,000 --> 00:44:57,840
I do want to play with Metas though.

622
00:44:57,840 --> 00:45:04,680
The most recent demo I saw, I think it was last week, had people speaking one language,

623
00:45:04,680 --> 00:45:10,120
but with very deliberate sort of emotional expression.

624
00:45:10,120 --> 00:45:17,360
So like being scared, being excited, whispering, and then translating that into another language

625
00:45:17,360 --> 00:45:24,120
spoken in that same person's voice, but maintaining the emotional performance.

626
00:45:24,120 --> 00:45:28,080
You just think about the use cases of this for call centers and all those other things

627
00:45:28,080 --> 00:45:29,080
that we've talked about.

628
00:45:29,080 --> 00:45:34,160
But as marketers being able to create audio versions of all your blog posts that don't

629
00:45:34,160 --> 00:45:40,800
sound sort of crappy and robotic like all voice synthesis has been, will there be an

630
00:45:40,800 --> 00:45:45,360
audio book production market like we know it?

631
00:45:45,360 --> 00:45:51,000
Or will you buy the text version on Kindle and you'll get three voices for free?

632
00:45:51,000 --> 00:45:56,880
But if you want the really cool voices or you know, what happens when famous orators

633
00:45:56,880 --> 00:46:02,520
who have lovely voices, you can now get all your books read for you by Morgan Freeman

634
00:46:02,520 --> 00:46:07,600
because they've licensed his voice out and you can get all your audio books read in his

635
00:46:07,600 --> 00:46:08,600
style.

636
00:46:08,600 --> 00:46:12,640
There's just so many things here that again, I think we're on a cusp of a revolution in

637
00:46:12,640 --> 00:46:18,520
terms of how we use spoken audio in so many different areas.

638
00:46:18,520 --> 00:46:22,960
And I suspect that the likes of Andres and Horowitz are looking at Eleven Labs as being

639
00:46:22,960 --> 00:46:27,780
major fuel for that, where they would think the commercial applications of this technology

640
00:46:27,780 --> 00:46:32,760
are so wide ranging that the revenue is massive that even if there's eight companies that

641
00:46:32,760 --> 00:46:38,040
all go to market with very popular products, they all still make an absolute ton of cash,

642
00:46:38,040 --> 00:46:39,960
which of course could be very possible as well.

643
00:46:39,960 --> 00:46:41,480
Let's talk website builders.

644
00:46:41,480 --> 00:46:42,480
You've been playing, Martin.

645
00:46:42,480 --> 00:46:44,440
Talk to us about how you feel about website builders.

646
00:46:44,440 --> 00:46:50,720
There's been a lot of hype around AI tools and the demo videos, as we've seen and have

647
00:46:50,720 --> 00:46:56,920
in fact talked about today, showing off products in their very best light.

648
00:46:56,920 --> 00:47:05,800
And one that always stuck in my mind was the Wix AI website builder, which looked so cool.

649
00:47:05,800 --> 00:47:11,760
The idea that you could enter your prompt and it would go away and build you this site

650
00:47:11,760 --> 00:47:19,640
and then you could edit the site by further prompting, saying make it more modern, make

651
00:47:19,640 --> 00:47:24,000
it more dynamic, make it more classic or what have you.

652
00:47:24,000 --> 00:47:26,200
And it would make all of these things on the fly.

653
00:47:26,200 --> 00:47:31,160
And the video, I remember at the time we talked about it, just looked so cool and I was really

654
00:47:31,160 --> 00:47:34,120
excited to get my hands on it.

655
00:47:34,120 --> 00:47:37,880
Well that now exists as a product.

656
00:47:37,880 --> 00:47:41,400
You can use it, the Wix AI website builder.

657
00:47:41,400 --> 00:47:50,800
And to say that it is not what I expected based on the video is something of an understatement.

658
00:47:50,800 --> 00:47:56,520
So the mechanism for this is you go on to the website builder and it starts off with

659
00:47:56,520 --> 00:48:00,440
a chat interface and it will say, tell us the name of your business.

660
00:48:00,440 --> 00:48:03,120
You type in the name of your business.

661
00:48:03,120 --> 00:48:07,340
And then it says, tell us about your business and the services and products that you sell.

662
00:48:07,340 --> 00:48:10,360
And then you type in that as well.

663
00:48:10,360 --> 00:48:11,840
And then there's a couple more questions.

664
00:48:11,840 --> 00:48:17,020
Effectively, what they've done is they've turned a form into a chat interface to make

665
00:48:17,020 --> 00:48:21,040
it AI because that's what people do, isn't it?

666
00:48:21,040 --> 00:48:22,400
They chat with chat GPT.

667
00:48:22,400 --> 00:48:24,840
So that's AI, isn't it?

668
00:48:24,840 --> 00:48:27,840
So they give you a form, you fill out the form.

669
00:48:27,840 --> 00:48:34,320
Then on the next screen, you choose the color and style that you want.

670
00:48:34,320 --> 00:48:36,500
But it's a limited range.

671
00:48:36,500 --> 00:48:39,220
So there'll be six options.

672
00:48:39,220 --> 00:48:43,520
And you might be like, oh, I want the green and white one, sans serif font.

673
00:48:43,520 --> 00:48:46,340
So you choose that one.

674
00:48:46,340 --> 00:48:56,600
And then it picks a template and uses the GPT-3 API.

675
00:48:56,600 --> 00:49:03,580
I'm assuming that's what it's using to populate this template with some text.

676
00:49:03,580 --> 00:49:12,000
The thing is, on the examples that I've seen, it's really lacking in depth.

677
00:49:12,000 --> 00:49:17,560
It seems to take none of the input text that you give it from the prompt and turn it into

678
00:49:17,560 --> 00:49:21,240
something relevant on the page.

679
00:49:21,240 --> 00:49:24,240
I think it's trash, to be perfectly honest.

680
00:49:24,240 --> 00:49:28,240
I think it's a really poor implementation of it.

681
00:49:28,240 --> 00:49:32,960
And the reason I wanted to talk about it is because actually, I tried a version of this

682
00:49:32,960 --> 00:49:36,880
with another CMS from a company called Doric.

683
00:49:36,880 --> 00:49:42,480
So Doric is based out of Bangladesh, it's a no-code CMS.

684
00:49:42,480 --> 00:49:49,720
In fact, the artificially intelligent marketing website is built and hosted on the Doric website

685
00:49:49,720 --> 00:49:50,720
builder.

686
00:49:50,720 --> 00:49:57,240
And they literally this week launched their AI website builder in beta.

687
00:49:57,240 --> 00:50:00,240
So it's private beta.

688
00:50:00,240 --> 00:50:01,960
And it's pretty good.

689
00:50:01,960 --> 00:50:05,120
First of all, it's much more like standard prompting.

690
00:50:05,120 --> 00:50:09,480
It says, what's the name of your business and then give us a description of your business

691
00:50:09,480 --> 00:50:12,200
in 1000 characters.

692
00:50:12,200 --> 00:50:16,160
And you give it that and then you hit generate.

693
00:50:16,160 --> 00:50:20,680
And then it chooses one of the templates for you, but it uses your description to choose

694
00:50:20,680 --> 00:50:23,640
the style for you.

695
00:50:23,640 --> 00:50:27,640
So it will, and I've played with this in a few different versions now, and it does a

696
00:50:27,640 --> 00:50:28,840
pretty good job.

697
00:50:28,840 --> 00:50:32,680
Now you can go in and edit the global styles if you don't like what it generates.

698
00:50:32,680 --> 00:50:39,440
But if you're using an AI generator for your website, chances are you're not somebody with

699
00:50:39,440 --> 00:50:40,520
a huge design team.

700
00:50:40,520 --> 00:50:42,940
You're probably a solo entrepreneur.

701
00:50:42,940 --> 00:50:47,320
Maybe you're a marketing consultant or a yoga teacher or whatever.

702
00:50:47,320 --> 00:50:51,080
You're not there building websites all the time with this detail.

703
00:50:51,080 --> 00:50:53,520
So it does a very good job of choosing the style.

704
00:50:53,520 --> 00:50:55,880
And then it literally will build the website.

705
00:50:55,880 --> 00:51:01,400
You can see it filling in all the text in all of the different spots as it goes.

706
00:51:01,400 --> 00:51:02,400
And it does a really good job.

707
00:51:02,400 --> 00:51:03,960
It makes some of the detail.

708
00:51:03,960 --> 00:51:10,360
You can tell in the back end, they've really thought about taking your input and applying

709
00:51:10,360 --> 00:51:14,440
that to things like the product and services, applying that to what this section should

710
00:51:14,440 --> 00:51:16,400
be versus this section.

711
00:51:16,400 --> 00:51:18,200
And you don't see that in the Wix version.

712
00:51:18,200 --> 00:51:23,480
The Wix version slaps in some sections that feel like placeholders.

713
00:51:23,480 --> 00:51:29,480
And it almost feels like the text in there is placeholder as opposed to the direct version

714
00:51:29,480 --> 00:51:33,280
where the text that it puts in feels informed by your prompt.

715
00:51:33,280 --> 00:51:35,200
It feels relevant.

716
00:51:35,200 --> 00:51:40,800
So I think for a really small startup based out in Bangladesh, they have done a fantastic

717
00:51:40,800 --> 00:51:42,280
first stab at this.

718
00:51:42,280 --> 00:51:49,080
And it gives me hope that a decent AI website builder will come to the fore.

719
00:51:49,080 --> 00:51:53,200
And I've also seen the HubSpot AI website builder.

720
00:51:53,200 --> 00:51:57,800
I saw it at inbound and I saw a live demo and that's more similar to the Wix version

721
00:51:57,800 --> 00:52:00,360
in terms of the implementation of it.

722
00:52:00,360 --> 00:52:08,480
It's more of a guided wizard, a WYSIWYG kind of, and what you end up with is a WYSIWYG

723
00:52:08,480 --> 00:52:12,200
drag and drop template, but to get there, you go through a wizard asking you lots of

724
00:52:12,200 --> 00:52:17,440
questions, asking you to fill in some data, pick the style and template that you like

725
00:52:17,440 --> 00:52:18,440
the look of.

726
00:52:18,440 --> 00:52:23,180
And then it creates it with images that it's found from stock libraries and text that it's

727
00:52:23,180 --> 00:52:27,160
generated with with ChatGPT 3.5.

728
00:52:27,160 --> 00:52:28,160
And that does a good job.

729
00:52:28,160 --> 00:52:30,360
That's a good starting point.

730
00:52:30,360 --> 00:52:37,840
But the most pure version of an AI builder that I've seen is the version from Doric.

731
00:52:37,840 --> 00:52:41,860
And I think they've done a very good job for saying it literally launched in beta just

732
00:52:41,860 --> 00:52:42,860
this week.

733
00:52:42,860 --> 00:52:47,200
It sounds like it's worth going and having a play with Doric if you are interested in

734
00:52:47,200 --> 00:52:49,880
these types of AI driven website builders.

735
00:52:49,880 --> 00:52:56,040
I think listening to you talk about that, it really makes me reflect on my biggest challenge

736
00:52:56,040 --> 00:53:00,240
at the moment is I think I've got AI hype cycle whiplash.

737
00:53:00,240 --> 00:53:01,640
I'm like, yes, this is amazing.

738
00:53:01,640 --> 00:53:02,640
Oh, this sucks.

739
00:53:02,640 --> 00:53:03,640
Yes, this is amazing.

740
00:53:03,640 --> 00:53:04,640
Oh, this sucks.

741
00:53:04,640 --> 00:53:06,080
Like I'm in the excited phase.

742
00:53:06,080 --> 00:53:11,000
I'm in the trough of disillusionment and I'm flipping back and forward in a single day,

743
00:53:11,000 --> 00:53:15,960
depending on if I'm looking at video or I'm using GPT-4 for an interesting use case or

744
00:53:15,960 --> 00:53:18,280
I'm looking at a website builder.

745
00:53:18,280 --> 00:53:24,160
I can flip between those states in a single domain like video or if I'm moving between

746
00:53:24,160 --> 00:53:25,160
domains.

747
00:53:25,160 --> 00:53:31,040
That is the biggest challenge at the moment in AI and using it for all these different

748
00:53:31,040 --> 00:53:34,720
use cases because we're seeing an explosion, right?

749
00:53:34,720 --> 00:53:40,000
Text generation, image generation, video generation, audio generation and then layered on top of

750
00:53:40,000 --> 00:53:46,080
that what I would call quasi-thinking agents that can do some brainstorming and thinking

751
00:53:46,080 --> 00:53:47,080
for you, right?

752
00:53:47,080 --> 00:53:52,040
Like turning a prompt into a web page template.

753
00:53:52,040 --> 00:53:56,480
That's a multi-step process that's beyond just getting an output.

754
00:53:56,480 --> 00:54:00,920
We're racing ahead in so many different areas all at the same time.

755
00:54:00,920 --> 00:54:02,600
None of them are in a straight line.

756
00:54:02,600 --> 00:54:08,840
I think as a general user, it's very hard to know where you can obtain value with these

757
00:54:08,840 --> 00:54:09,920
different systems.

758
00:54:09,920 --> 00:54:16,160
I think if you go into using them and then you get a bad first output or it doesn't

759
00:54:16,160 --> 00:54:21,200
really generate the thing that you want, you could see why people quite quickly go, AI

760
00:54:21,200 --> 00:54:26,400
is fun and it's nice but it's not there yet because it doesn't give me the things I want.

761
00:54:26,400 --> 00:54:30,480
I suspect that a lot of people are actually running into that and then thinking and putting

762
00:54:30,480 --> 00:54:36,360
AI to one side for a bit and going, yeah, I can't do the things I need it to do.

763
00:54:36,360 --> 00:54:40,680
Listening to you talk about website builders nicely encapsulates that whiplash for me and

764
00:54:40,680 --> 00:54:42,840
it's definitely something I'm feeling.

765
00:54:42,840 --> 00:54:51,560
If I was going to give any tip to a marketer that is looking to maybe deploy AI in their

766
00:54:51,560 --> 00:54:57,600
product and announce it, maybe trail it with sneak preview type of thing, one way that

767
00:54:57,600 --> 00:55:03,680
you can help people like me and Paul avoid such a whiplash is to just manage expectations.

768
00:55:03,680 --> 00:55:08,280
Don't put out videos showcasing stuff that it can't actually do.

769
00:55:08,280 --> 00:55:12,800
That's not what we need because we're just going to be disappointed when we try and use

770
00:55:12,800 --> 00:55:15,320
it and that's not what you want in your users.

771
00:55:15,320 --> 00:55:16,320
Right.

772
00:55:16,320 --> 00:55:18,000
We're drifting in on time, Martin.

773
00:55:18,000 --> 00:55:22,840
We probably need to swing through these last few stories quite quickly.

774
00:55:22,840 --> 00:55:25,240
Let's briefly touch on this MIT study.

775
00:55:25,240 --> 00:55:30,440
There was a recent study by MIT's Computer Science and Artificial Intelligence Laboratory

776
00:55:30,440 --> 00:55:32,960
with MIT Sloan.

777
00:55:32,960 --> 00:55:41,720
Basically this study looked into the potential impact of AI on jobs with a focus on computer

778
00:55:41,720 --> 00:55:43,680
vision.

779
00:55:43,680 --> 00:55:49,240
The core finding from the study is that only about 23% of tasks that require vision are

780
00:55:49,240 --> 00:55:53,640
currently economically viable to automate using AI.

781
00:55:53,640 --> 00:55:59,400
This is probably one of those research reports that is a counterbalance to other search reports

782
00:55:59,400 --> 00:56:04,480
that are saying AI is coming for everybody's jobs and in one year nobody will have a job

783
00:56:04,480 --> 00:56:07,440
because AI will do everything.

784
00:56:07,440 --> 00:56:14,920
What this study is saying is AI's potential is large, especially on giving computers vision

785
00:56:14,920 --> 00:56:21,440
to see things and do things, but the sheer cost of doing that massively outweighs having

786
00:56:21,440 --> 00:56:24,040
humans do it.

787
00:56:24,040 --> 00:56:28,660
I think this study was quite a nice counterbalance to some of those other studies and it further

788
00:56:28,660 --> 00:56:33,800
shines the light on the cost of using AI.

789
00:56:33,800 --> 00:56:40,280
There was another story this week about just how much energy consumption is used to run

790
00:56:40,280 --> 00:56:41,280
these AI models.

791
00:56:41,280 --> 00:56:45,400
I'm going to get this quote wrong, but there was a quote, I think it was in the Wall Street

792
00:56:45,400 --> 00:56:51,760
Journal that was something of the equivalent of asking a simple question of chat GPT is

793
00:56:51,760 --> 00:56:57,000
like driving a Lamborghini to do pizza deliveries.

794
00:56:57,000 --> 00:57:01,120
I think that's probably true as well.

795
00:57:01,120 --> 00:57:04,760
You see Sam Altman talk a lot about the next revolution we're going to require is going

796
00:57:04,760 --> 00:57:10,520
to be energy generation because the energy that we're going to need to run all of these

797
00:57:10,520 --> 00:57:18,620
server farms and all of these graphics processing units or these emerging chip architectures

798
00:57:18,620 --> 00:57:23,880
to drive large language models, it's all going to take a lot of energy, a lot of development

799
00:57:23,880 --> 00:57:25,800
of other technologies.

800
00:57:25,800 --> 00:57:32,880
So, as folks, as humans, we've got a little bit more time maybe before AI comes and takes

801
00:57:32,880 --> 00:57:34,280
on our jobs than we thought.

802
00:57:34,280 --> 00:57:39,520
Yeah, that's an interesting point from Sam Altman, particularly as he has his fingers

803
00:57:39,520 --> 00:57:46,000
in the energy pie, doesn't he, with Helion, the nuclear energy startup.

804
00:57:46,000 --> 00:57:52,280
That's no doubt going to be powering those server farms and the rest of us for years

805
00:57:52,280 --> 00:57:53,280
to come.

806
00:57:53,280 --> 00:57:58,480
Yeah, if you look inside baseball on this one, the fact that Sam Altman, you see these

807
00:57:58,480 --> 00:58:03,480
different things emerge where he's involved in companies designing new chips, right, because

808
00:58:03,480 --> 00:58:06,960
he's clearly going to see how many chips the world's going to need if everything's powered

809
00:58:06,960 --> 00:58:07,960
by AI.

810
00:58:07,960 --> 00:58:10,960
He's looking at energy companies because he's like, I know how much energy is going to be

811
00:58:10,960 --> 00:58:14,520
required to run those chips.

812
00:58:14,520 --> 00:58:19,920
If you're inside the system on this, you can probably make some pretty healthy bets on

813
00:58:19,920 --> 00:58:27,040
where the economic drivers are going to be that enable this AI revolution.

814
00:58:27,040 --> 00:58:29,160
So it doesn't surprise me.

815
00:58:29,160 --> 00:58:32,680
In fact, it's probably worth watching what a lot of these people are investing their

816
00:58:32,680 --> 00:58:38,120
time and their money in outside of AI because they probably know things that we all don't.

817
00:58:38,120 --> 00:58:41,720
Not that I'm qualified in any way to give investment advice, and I'm not, so please

818
00:58:41,720 --> 00:58:46,040
don't take my advice, but I just think it's interesting.

819
00:58:46,040 --> 00:58:50,560
Let's talk mobile phones, Martin, because we've been talking about one of the cool

820
00:58:50,560 --> 00:58:54,320
things that we probably will see over time is that small models can be run natively on

821
00:58:54,320 --> 00:58:58,480
a phone without having to go up to the internet and servers, which obviously has a range of

822
00:58:58,480 --> 00:59:00,520
benefits that we've talked about.

823
00:59:00,520 --> 00:59:07,600
But Samsung Galaxy S24 has been announced and released over the last week or so, and

824
00:59:07,600 --> 00:59:14,320
they're, to my knowledge, the first big phone manufacturer to lean into AI in a big way.

825
00:59:14,320 --> 00:59:20,200
So what do we have that the S24 is going to be bringing to the market?

826
00:59:20,200 --> 00:59:25,160
It's really an AI forward mobile device.

827
00:59:25,160 --> 00:59:30,240
They've really leaned into the messaging and basically the chipsets and the technology

828
00:59:30,240 --> 00:59:35,080
independent allows there to be a lot of AI on device.

829
00:59:35,080 --> 00:59:43,520
So they talk about live translate and an interpreter offering real-time voice and text translations,

830
00:59:43,520 --> 00:59:46,320
which on the demo looked very cool.

831
00:59:46,320 --> 00:59:50,000
They've got the Pro Visual Engine.

832
00:59:50,000 --> 00:59:58,320
So this is a comprehensive suite of AI powered tools for image capturing and editing.

833
00:59:58,320 --> 01:00:06,580
So things like enhanced night photography, high quality video editing and photography

834
01:00:06,580 --> 01:00:10,020
manipulation on device.

835
01:00:10,020 --> 01:00:14,640
We've got this intuitive search with circle to search.

836
01:00:14,640 --> 01:00:23,320
So this is a new gesture driven feature, which allows Google to search by simply drawing

837
01:00:23,320 --> 01:00:26,440
a circle and highlighting something on the screen.

838
01:00:26,440 --> 01:00:30,800
So it will do a kind of Google image search or something similar with that, which is a

839
01:00:30,800 --> 01:00:32,560
great new input.

840
01:00:32,560 --> 01:00:37,240
I always like these new interfaces where you get a new way of interacting with tech.

841
01:00:37,240 --> 01:00:38,960
I'm like, that's pretty cool.

842
01:00:38,960 --> 01:00:41,160
That one is cool.

843
01:00:41,160 --> 01:00:48,200
Because if you think about, for marketers, we talk about how if we speak to computers

844
01:00:48,200 --> 01:00:53,280
most of the time, instead of typing and reading stuff, I'm not necessarily saying it's all

845
01:00:53,280 --> 01:01:00,720
or nothing on that, but as user behavior changes there, it just puts more pressure on search

846
01:01:00,720 --> 01:01:03,520
engine optimization as a strategy.

847
01:01:03,520 --> 01:01:07,880
So many things we've talked about on the podcast in terms of Google's generative search experience

848
01:01:07,880 --> 01:01:11,600
and it's showing more and more answers in the search pane and not driving more traffic

849
01:01:11,600 --> 01:01:12,680
to your website.

850
01:01:12,680 --> 01:01:17,680
Then we've got people using what we expect is imminent in terms of smarter assistance

851
01:01:17,680 --> 01:01:19,240
that we can ask questions of that.

852
01:01:19,240 --> 01:01:21,280
Again, we're not looking at a search results page.

853
01:01:21,280 --> 01:01:23,660
We're not clicking through to other websites.

854
01:01:23,660 --> 01:01:29,440
Now we've got circle things in images that you think are cool and then Google will search

855
01:01:29,440 --> 01:01:32,440
for content around that.

856
01:01:32,440 --> 01:01:37,160
The potential complexity and challenges ahead for SEO professionals over the next two or

857
01:01:37,160 --> 01:01:44,120
three years, I don't think should be underestimated as people's engagement mechanisms with computers

858
01:01:44,120 --> 01:01:49,960
and phones evolves to be drawing circles and speaking to agents.

859
01:01:49,960 --> 01:01:55,560
Yeah, SEO is, I'm really fascinated to see how quickly that world changes.

860
01:01:55,560 --> 01:01:58,040
And it could still be slow, I don't know, right?

861
01:01:58,040 --> 01:01:59,720
But that fascinates me.

862
01:01:59,720 --> 01:02:09,200
Yeah, just on the idea of it being one of the first from an AI on device manufacturers,

863
01:02:09,200 --> 01:02:17,520
the Google Pixel 8 and the Pixel 8 Pro, they have a lot of AI baked into them as well.

864
01:02:17,520 --> 01:02:21,040
I don't think as much as this particular device.

865
01:02:21,040 --> 01:02:26,480
OnePlus, which is the manufacturer of device that I use, they just announced their OnePlus

866
01:02:26,480 --> 01:02:35,480
12 device this week and Qualcomm were there at the keynote talking about the on device

867
01:02:35,480 --> 01:02:37,440
AI capabilities of that phone.

868
01:02:37,440 --> 01:02:44,640
Again, don't think it quite matches that of the Galaxy S24 series, but we're seeing more

869
01:02:44,640 --> 01:02:47,440
and more of this coming on device.

870
01:02:47,440 --> 01:02:53,280
So we're probably going to see most flagship devices launched over the next year being

871
01:02:53,280 --> 01:02:59,680
capable of running, I think, 10 billion parameter models on device.

872
01:02:59,680 --> 01:03:01,280
Yeah, that's interesting.

873
01:03:01,280 --> 01:03:04,800
And ultimately, we don't want to give them wrong impression here on the podcast.

874
01:03:04,800 --> 01:03:09,480
AI has been underpinning a number of things in phones for years, right?

875
01:03:09,480 --> 01:03:14,120
If you use Google Photos and it's got smart tagging capabilities and you can type in the

876
01:03:14,120 --> 01:03:18,200
word dog into your search bar and it will go find all your pictures with dogs in.

877
01:03:18,200 --> 01:03:25,280
So it's less that phones have got AI now, I think phones always had AI or have had forms

878
01:03:25,280 --> 01:03:27,080
of AI for many years.

879
01:03:27,080 --> 01:03:33,680
I think what's interesting here is models on device and everybody getting on the AI hype

880
01:03:33,680 --> 01:03:40,080
train and making AI a differentiator for their platform, which will very soon not be a differentiator

881
01:03:40,080 --> 01:03:41,080
probably.

882
01:03:41,080 --> 01:03:45,040
I think one of the things I thought was cool here, and I don't know how it will play out,

883
01:03:45,040 --> 01:03:50,360
is some of these onboard, assuming they're onboard, transcription and translation services.

884
01:03:50,360 --> 01:03:55,040
Because one of my use cases back in the day for Otter, which is a transcription, one of

885
01:03:55,040 --> 01:04:00,680
the first transcription services, was pre pandemic having my phone, putting it on the

886
01:04:00,680 --> 01:04:05,280
table for a meeting and then having Otter transcribe it for me.

887
01:04:05,280 --> 01:04:10,720
But if a lot of that stuff ends up just onboard my phone as it is, am I going to stop using

888
01:04:10,720 --> 01:04:18,600
certain platforms like Otter for creating transcriptions of calls, for example.

889
01:04:18,600 --> 01:04:24,120
And I think that transcription eMarkets in extreme dynamic flux, right, because it's

890
01:04:24,120 --> 01:04:29,240
being baked into Teams and Zoom and Google.

891
01:04:29,240 --> 01:04:34,240
I think if I was CEO of Otter and similar tools, I'd be a bit worried honestly, because

892
01:04:34,240 --> 01:04:36,400
it's just being baked into everything.

893
01:04:36,400 --> 01:04:37,400
Yeah.

894
01:04:37,400 --> 01:04:39,320
So I use a tool called Laxis.

895
01:04:39,320 --> 01:04:43,000
We've had Eric from Laxis on the podcast recently.

896
01:04:43,000 --> 01:04:49,000
I use this in Google Meet all the time and Google Meet has recently launched its own

897
01:04:49,000 --> 01:04:51,320
transcription tool that you can have turned on.

898
01:04:51,320 --> 01:04:56,120
And if you join a meeting where it's turned on, it tells everybody that it's being transcribed.

899
01:04:56,120 --> 01:05:01,360
I still use Laxis over, even though I'm in the Google workspace ecosystem, I still use

900
01:05:01,360 --> 01:05:02,560
Laxis.

901
01:05:02,560 --> 01:05:07,160
And the reason for that is Laxis can join calls in other services.

902
01:05:07,160 --> 01:05:12,240
So by having it connected to my calendar, if it's my Google Meet, it just transcribes

903
01:05:12,240 --> 01:05:13,240
it in the background.

904
01:05:13,240 --> 01:05:18,840
But if I get invited to a Teams call for a client or something like that, I still get

905
01:05:18,840 --> 01:05:22,160
my own transcription because Laxis joins that call.

906
01:05:22,160 --> 01:05:27,000
So the idea of it being cross-platform is really critical for me.

907
01:05:27,000 --> 01:05:28,000
Interesting.

908
01:05:28,000 --> 01:05:30,360
Yeah, I think that makes a lot of sense.

909
01:05:30,360 --> 01:05:35,400
Right, two more stories listeners and then we'll let you to go about your day.

910
01:05:35,400 --> 01:05:42,200
The first one very quickly is an interesting story where AI was being used for political

911
01:05:42,200 --> 01:05:43,200
misinformation.

912
01:05:43,200 --> 01:05:44,200
So you've probably seen this in the news.

913
01:05:44,200 --> 01:05:50,200
It's been fairly high profile, but basically a robo-caller using a fake AI generated voice

914
01:05:50,200 --> 01:05:56,700
of President Joe Biden urging Democrats to stay home and not participate in primary elections.

915
01:05:56,700 --> 01:06:01,680
And we talked about this a bit on the podcast and we talked about 11 Labs today.

916
01:06:01,680 --> 01:06:10,200
The ability of AI to clone and mimic people's voices and with tools like HeyGen, even video,

917
01:06:10,200 --> 01:06:13,800
we're drifting ever closer to that world where you can't trust what you see.

918
01:06:13,800 --> 01:06:15,240
You can't trust a mid-journey image.

919
01:06:15,240 --> 01:06:16,740
You can't trust a voice.

920
01:06:16,740 --> 01:06:18,240
You can't trust a video.

921
01:06:18,240 --> 01:06:23,440
I know people who are already talking about having like safe words or stories that they

922
01:06:23,440 --> 01:06:28,880
can share with each other to prove that it's actually them that's calling, right?

923
01:06:28,880 --> 01:06:33,480
Like your grandmother rings you to say that she's got herself in some trouble and she

924
01:06:33,480 --> 01:06:38,280
needs you to send some money and it sounds exactly like her, but actually it's synthetic

925
01:06:38,280 --> 01:06:40,640
voice for example.

926
01:06:40,640 --> 01:06:46,000
And this is the first sort of mainstream media example of it being used to influence pretty

927
01:06:46,000 --> 01:06:48,960
major processes like the US election cycle.

928
01:06:48,960 --> 01:06:56,060
So we won't go into too much detail on this, but I think as marketers and communications

929
01:06:56,060 --> 01:07:02,160
professionals we will need to work hard to set the standard to build trust in content

930
01:07:02,160 --> 01:07:08,600
that's supported with AI tools because the amount of mistrust that's going to get generated

931
01:07:08,600 --> 01:07:11,680
by bad actors is going to be large enough.

932
01:07:11,680 --> 01:07:16,820
So labeling things as being generated by AI or supported with AI I think is going to become

933
01:07:16,820 --> 01:07:22,560
more and more important as the tools' abilities to pull the wool over our eyes only increase.

934
01:07:22,560 --> 01:07:26,280
And so it's probably something that we all as humans need to be aware of in terms of

935
01:07:26,280 --> 01:07:30,880
fact checking and really being sure that we can trust our sources, but as marketers maybe

936
01:07:30,880 --> 01:07:36,540
a bit of extra requirements upon us to do the right thing in these areas.

937
01:07:36,540 --> 01:07:41,880
As we see more and more examples of it, and today there was the example of Taylor Swift

938
01:07:41,880 --> 01:07:48,920
with some deep fake pornography shared widely on Twitter and she's come out and said this

939
01:07:48,920 --> 01:07:51,920
is terrible and it's got lots of headlines.

940
01:07:51,920 --> 01:07:55,880
As we see more and more examples of this, I think people are going to get more and more

941
01:07:55,880 --> 01:07:57,000
aware of it.

942
01:07:57,000 --> 01:08:03,000
When it's high profile cases like Joe Biden, like Taylor Swift who let's be honest is the

943
01:08:03,000 --> 01:08:11,160
biggest celebrity in music right now, people are going to get smarter and more discerning.

944
01:08:11,160 --> 01:08:17,080
I like to think, I have trust in humanity where people will just take a second and say,

945
01:08:17,080 --> 01:08:20,360
hold on, is this real?

946
01:08:20,360 --> 01:08:23,560
I think you need to be worried Martin because for me you're the biggest celebrity in the

947
01:08:23,560 --> 01:08:29,760
world of AI and marketing and so having someone clone your voice from this raw source podcast

948
01:08:29,760 --> 01:08:31,840
here I think is a major risk for you.

949
01:08:31,840 --> 01:08:36,200
Yeah and there's enough video material out there that they might turn it into some deep

950
01:08:36,200 --> 01:08:41,000
fake porn as well so strap in everyone.

951
01:08:41,000 --> 01:08:48,320
I'm glad you didn't mispronounce your words there but yes, usually most podcast hosts

952
01:08:48,320 --> 01:08:51,640
would edit this bit out but we're not going to do that.

953
01:08:51,640 --> 01:08:53,520
Let's move on to our last story Martin.

954
01:08:53,520 --> 01:08:56,240
Let's go with this Microsoft Copilot.

955
01:08:56,240 --> 01:09:01,440
We talked a bit about ChatGPT for Teams and how finally small businesses can tap into

956
01:09:01,440 --> 01:09:05,000
the commercial power of ChatGPT without having to worry about the information they put into

957
01:09:05,000 --> 01:09:06,360
ChatGPT.

958
01:09:06,360 --> 01:09:09,960
Hot off the heels, pretty much within a day or two.

959
01:09:09,960 --> 01:09:10,960
Microsoft Copilot.

960
01:09:10,960 --> 01:09:12,720
What happened here?

961
01:09:12,720 --> 01:09:20,080
They have released Microsoft Copilot to everybody so it was previously restricted to organizations

962
01:09:20,080 --> 01:09:24,600
that had, was it 300 seats I think you needed to buy.

963
01:09:24,600 --> 01:09:30,200
Whereas now they're saying organizations of any size can sign up maybe a minimum of two

964
01:09:30,200 --> 01:09:33,580
seats for the business version, I'm not entirely sure.

965
01:09:33,580 --> 01:09:39,960
But even the home and family edition of Microsoft 365 has a Copilot upgrade as well which means

966
01:09:39,960 --> 01:09:44,880
that you now have access to Copilot Chat which was previously Bing Chat.

967
01:09:44,880 --> 01:09:52,760
You have Copilot in Outlook which does summarization of emails as well as drafting replies for

968
01:09:52,760 --> 01:09:53,760
you.

969
01:09:53,760 --> 01:10:00,800
Copilot for PowerPoint which is giving you a starting point if you just write a bit of

970
01:10:00,800 --> 01:10:05,400
a prompt it will create the PowerPoint deck for you.

971
01:10:05,400 --> 01:10:11,480
And some really cool demos of that actually and it does a pretty good job, a kind of B

972
01:10:11,480 --> 01:10:14,200
plus level performance.

973
01:10:14,200 --> 01:10:22,320
I'm not going to knock it like I did when we reviewed the Google Duets, Google Slides

974
01:10:22,320 --> 01:10:24,480
version.

975
01:10:24,480 --> 01:10:28,080
It seems to do a good job when you provide it with like a Word document.

976
01:10:28,080 --> 01:10:31,440
So like say you've got a report that you've written and you need to turn it into a presentation

977
01:10:31,440 --> 01:10:34,840
it seems to do quite a good job of doing that conversion.

978
01:10:34,840 --> 01:10:38,600
So I think that's going to save quite a lot of people time actually to get that starter

979
01:10:38,600 --> 01:10:44,280
for 10 in terms of getting it all the material put into a PowerPoint style laid out and all

980
01:10:44,280 --> 01:10:45,280
that stuff.

981
01:10:45,280 --> 01:10:50,300
They've also got Copilot in Word which is pretty much bringing content generation into

982
01:10:50,300 --> 01:10:56,200
Microsoft Word and Copilot in Excel which at the moment is quite limited and not really

983
01:10:56,200 --> 01:10:58,200
worth talking about.

984
01:10:58,200 --> 01:11:04,760
It's the area that I have the most hope for because I think when that gets good we'll

985
01:11:04,760 --> 01:11:14,880
start to see some very exciting data analysis functionalities put into the hands of everybody.

986
01:11:14,880 --> 01:11:21,040
Pricing of this is $20 per month per user and is available now.

987
01:11:21,040 --> 01:11:23,080
Yeah, it's a fascinating one.

988
01:11:23,080 --> 01:11:28,040
I reflect back on us getting super excited about the Microsoft Copilot like launch video

989
01:11:28,040 --> 01:11:30,880
which was about last March, April time.

990
01:11:30,880 --> 01:11:37,840
So it's taken the best part of a year for this to be in the hands of everyone which

991
01:11:37,840 --> 01:11:43,360
I think relatively speaking for what they've had to build and deploy and test it through

992
01:11:43,360 --> 01:11:49,160
the larger organizations, the enterprises that have had access is probably fair game

993
01:11:49,160 --> 01:11:54,360
but it's not quite as powerful as the demo made it look and we still don't have the

994
01:11:54,360 --> 01:11:55,360
Excel part.

995
01:11:55,360 --> 01:11:59,260
So it comes back to an earlier part of our conversation, Martin, in terms of if you see

996
01:11:59,260 --> 01:12:05,600
something cool in a research paper or a demo, how long until the power of that is in your

997
01:12:05,600 --> 01:12:06,600
hands?

998
01:12:06,600 --> 01:12:12,280
Probably 6 to 12 months, probably closer to 12 and reduce your expectations on that tool

999
01:12:12,280 --> 01:12:15,600
by 25 to 50% based on what you saw in the demo.

1000
01:12:15,600 --> 01:12:20,040
Seems to be a rule of thumb in terms of the commercial application of a lot of these AI

1001
01:12:20,040 --> 01:12:22,640
based tools, especially the generative ones.

1002
01:12:22,640 --> 01:12:25,120
Right, I think we'll call it there.

1003
01:12:25,120 --> 01:12:26,120
Thank you for sticking with us folks.

1004
01:12:26,120 --> 01:12:27,920
There was a lot to cover this week.

1005
01:12:27,920 --> 01:12:30,920
Thank you to you, Martin, as always for your time and insights.

1006
01:12:30,920 --> 01:12:31,920
Much appreciated.

1007
01:12:31,920 --> 01:12:39,800
Well, thank you and I hope you can just have a moment to sit and think about what Jürgen

1008
01:12:39,800 --> 01:12:43,360
Klopp has done for Liverpool and be thankful that it happened.

1009
01:12:43,360 --> 01:12:47,840
You know, I managed to forget that over the course of the last hour and now you've just

1010
01:12:47,840 --> 01:12:49,600
brought it all back for me.

1011
01:12:49,600 --> 01:12:51,120
Sorry about that.

1012
01:12:51,120 --> 01:12:57,000
It's always 5pm somewhere so it's probably time for a beer and a tear.

1013
01:12:57,000 --> 01:13:01,000
But yes, that's now going to take up the focus for me for the rest of the day.

1014
01:13:01,000 --> 01:13:02,880
I'm sure I've got nothing better to do.

1015
01:13:02,880 --> 01:13:03,880
Yes I do.

1016
01:13:03,880 --> 01:13:05,440
Right, listeners, thank you for your time.

1017
01:13:05,440 --> 01:13:08,920
We'll look forward to giving you the news in our next episode.

1018
01:13:08,920 --> 01:13:27,920
Until then, bye everyone.